Two-dimensional medical image segmentation network and method based on U-shaped structure

Abstract

The application discloses a two-dimensional medical image segmentation network and method based on a U-shaped structure, which inputs a medical image into a feature encoder module, uses a pre-trained ResNet-18 to obtain rich and complex features of the input image, and accelerates the convergence speed of the network; the feature map extracted by the feature encoder module is input into a multi-receptive field extraction module to extract semantic information of different receptive fields, and high-level feature maps are generated by using context information; the high-level feature maps enter a feature decoder module after passing through an SE module; wherein, the SE module allocates different weights to different channels according to different contributions of the channels to the segmentation task, so that the network focuses on specific parts of the input; and the feature decoder outputs a final prediction segmentation map after decoding. The application effectively improves the segmentation effect of the two-dimensional medical image.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of neural networks, in particular to a two-dimensional medical image segmentation network and method based on a U-shaped structure. BACKGROUND

[0002] A medical image segmentation method and system based on deep learning in the prior art. In the down-sampling process, two layers of adjacent feature layers with different resolutions in any of the historical magnetic resonance imaging (MRI) modality images in the training set are input into a multi-level feature re-extraction and aggregation neural network model to determine a segmented MRI modality image. The two layers of adjacent feature layers with different resolutions include a low-resolution feature layer and a high-resolution feature layer; the two layers of adjacent feature layers with different resolutions sequentially pass through a residual convolution unit, a resolution fusion unit and an aggregation unit to determine the segmented MRI modality image. In the down-sampling process, the method causes loss of spatial information due to continuous pooling operations.

[0003] A medical image automatic segmentation method based on multi-path attention fusion, in which an image is input into a multi-path encoder, four paths composed of 1, 2, 3 and 4 residual blocks respectively, then the outputs of each path are fused through an attention mechanism, and then a decoder is used to decode and output a segmentation image. Although the method uses attention to fuse the features output by different paths, it lacks the use of context information and global information.

[0004] The existing model causes loss of spatial information due to pooling operations in the down-sampling process. Context information and global information are less utilized. The pixel class imbalance problem in medical images causes the model to pay more attention to uninteresting areas during the training process, resulting in poor segmentation effect of the model. SUMMARY

[0005] The application aims to provide a two-dimensional medical image segmentation network and method based on a U-shaped structure.

[0006] The technical solution adopted by the application is:

[0007] The two-dimensional medical image segmentation network based on the U-shaped structure comprises the following modules:

[0008] The feature encoder module: a pre-trained ResNet-18 is used as an encoder to increase the depth of the network, and a shortcut mechanism is used in the branch to avoid gradient disappearance and accelerate network convergence. Max-pooling and stride convolution are used for down-sampling to reduce the size of the image.

[0009] Multi-receptive field extraction module: used to extract semantic information of different receptive fields and generate more high-level feature maps, integrated encoder-decoder structure, the output feature map of the encoding part can capture context information, and strengthen the use of global information;

[0010] Feature decoder module: decode the obtained high-level feature map to restore the size of the initial image; the Squeeze-and-Excitation (SE) module used assigns corresponding weights to different channels for different contributions to the segmentation task, so that the network focuses on specific input parts.

[0011] Further, the feature encoder module uses stride convolution with a step size of 2 to reduce the image size at the beginning, then uses maximum pooling to further reduce the image size, and then uses stride convolution with a step size of 2 to reduce the image size. After two convolutions, the input feature map becomes the output feature map; the adjusted input feature map and the output feature map are subjected to element-wise addition operation, increasing the amount of feature information in each dimension, and finally passing through the ReLU activation function to obtain the final output; the expression formula of the feature encoder module is as follows:

[0012]

[0013] wherein, Residual Basicblock, n represents the number of repetitions, i represents the number, and MaxPool(·) represents maximum pooling, Conv(7x7) represents a convolution operation with a convolution kernel size of 7x7.

[0014] Further, the multi-receptive field extraction module includes 4 cascaded branches, wherein a single 3x3 convolution branch and an average pooling Avgpooling branch are used to extract different feature information under the same receptive field, and the corresponding receptive field is 3; the other two branches are used to extract features of different scales, one branch uses two 3x3 convolutions to extract features; the other branch uses dilated convolution with dilation rates of 1, 2 and 3 to increase the receptive field of convolution without introducing additional parameters, and is applied to each pixel point in the receptive field; the feature maps obtained by the 4 cascaded branches are added to the input feature map;

[0015] The expression formula of the multi-receptive field extraction module is as follows:

[0016]

[0017] output (·) output represents the output feature map, x represents the input feature map, F i (·) represents the output of the i-th branch.

[0018] Further, the feature decoder module has 5 layers, each of the first 4 layers includes 1 Deconvolution TransposedConvolution, 1 Squeeze-and-Excitation (SE) module and 2 3x3 convolutions; the 5th layer includes 1 deconvolution, 1 3x3 convolution and 1 1x1 convolution; the feature map is input into the SE module, and a channel weight is obtained after global average pooling and a full connection layer, and a new feature map is obtained by multiplying the channel weight and the feature map.

[0019] The two-dimensional medical image segmentation method based on the U-shaped structure comprises the following steps:

[0020] Step S1: input the medical image into the feature encoder module, use the pre-trained ResNet-18 to obtain rich and complex features of the input image, and accelerate the convergence speed of the network;

[0021] Step S2: input the feature map extracted by the feature encoder module into the multi-receptive field extraction module to extract semantic information of different receptive fields, and generate high-level feature maps using context information;

[0022] Step S3: the high-level feature map is input into the feature decoder module after Squeeze-and-Excitation (SE); wherein Squeeze-and-Excitation (SE) allocates different weights to different channels according to their different contributions to the segmentation task, so that the network focuses on specific parts of the input;

[0023] Step S4: the feature decoder outputs the final prediction segmentation map after decoding.

[0024] Further, step S1 comprises the following steps:

[0025] Step S1-1: input the medical image through a convolution kernel size of 7x7 and a stride of 2 to reduce the image size by half and the channel number to 64;

[0026] Step S1-2: the output feature map of step S1-1 is subjected to maximum pooling and 2 layers of Residual BasicBlock to reduce the image size by half and the channel number to 64;

[0027] Step S1-3: the output feature map of step S1-2 is subjected to 2 layers of Residual BasicBlock to reduce the image size by half and the channel number to 128;

[0028] Step S1-4: the output feature map of step S1-3 is subjected to 2 layers of Residual BasicBlock to reduce the image size by half and the channel number to 256;

[0029] Step S1-5: The output feature map of step S1-4 is input into two layers of Residual BasicBlock to obtain a feature map with an image size halved and a channel number of 512.

[0030] Further, step S2 includes the following steps:

[0031] Step S2-1: The extracted feature map is input into a convolutional layer with a kernel size of 3x3, a BN layer, and a ReLU activation function to obtain a corresponding output.

[0032] Step S2-2: The extracted feature map is input into a convolutional layer with a kernel size of 1x1 and two convolutional layers with a kernel size of 3x3 to obtain a corresponding output.

[0033] Step S2-3: The extracted feature map is input into three dilated convolutional layers with a kernel size of 3x3 and dilation rates of 1, 2, and 3, respectively, to obtain a corresponding output.

[0034] Step S2-4: The extracted feature map is input into an average pooling layer with a kernel size of 3x3 and a convolutional layer with a kernel size of 1x1 to obtain a corresponding output.

[0035] Step S2-5: The extracted feature map is element-wise added to the outputs of steps S2-1, S2-2, S2-3, and S2-4 to obtain a high-level feature map.

[0036] Further, step S3 includes the following steps:

[0037] Step S3-1: An adaptive mapping is learned through deconvolution to restore features with more detailed information. The feature map is upsampled by a factor of two in both width and height through a deconvolution operation.

[0038] Step S3-2: The upsampled feature map is connected to the output of the corresponding convolutional layer of the encoder through a skip connection.

[0039] Step S3-3: The connected feature map is input into a Squeeze-and-Excitation (SE) module. A channel weight is obtained after global average pooling and a fully connected layer. The channel weight is multiplied by the connected feature map to obtain a new feature map.

[0040] Step S3-4: The new feature map is then subjected to two convolutions.

[0041] Step S3-5: Steps S3-1, S3-2, S3-3, and S3-4 are repeated four times in sequence.

[0042] The application adopts the above technical scheme, and the RAAU-Net model is composed of a feature encoder module, a multi-receptive field extraction module and a feature decoder module, wherein the multi-receptive field extraction module is proposed by the application and is composed of an Inception-ASPP-Residual (IAR) module. The module extracts semantic information of different receptive fields from the feature map output by the encoding part and generates more high-level feature maps. The application uses a pre-trained ResNet-18 to extract more rich and complex features, the multi-receptive field extraction module is used for encoder-decoder structure integration, the feature decoder module uses a Squeeze-and-Excitation (SE) attention mechanism to improve decoding capability, and the whole model presents a U-shaped structure.

[0043] The application alleviates the loss of spatial information caused by the pooling operation by using the residual network module, further improves the feature extraction capability of the network, extracts semantic information of different receptive fields through a multi-receptive field extraction module, thereby capturing context information, strengthening the use of global information and strengthening the feature expression capability of the network. The Squeeze-and-Excitation (SE) attention mechanism is used to improve the decoding capability of the model, so that the model pays attention to the region of interest during the training process, and improves the segmentation effect of the model. BRIEF DESCRIPTION OF DRAWINGS

[0044] The application will be further described in detail below in combination with the drawings and specific embodiments.

[0045] Figure 1 FIG. 1 is a schematic diagram of a two-dimensional medical image segmentation network structure based on a U-shaped structure according to the application;

[0046] Figure 2 FIG. 4 is a schematic diagram of a Residual Basicblock structure;

[0047] Figure 3 FIG. 6 is a schematic diagram of a multi-receptive field extraction module IAR structure;

[0048] Figure 4 FIG. 8 is a schematic diagram of an SE module structure;

[0049] Figure 5 FIG. 11 is a schematic diagram of a qualitative result comparison of the application and other algorithm models in retinal nerve segmentation using a DRIVE data set. EMBODIMENT

[0050] In order to make the purpose, technical scheme and advantages of the embodiments of the application more clear, the technical scheme of the embodiments of the application will be described clearly and completely below in combination with the drawings in the embodiments of the application.

[0051] As Figures 1 to 5As shown in the first aspect of the present application, the present application discloses a two-dimensional medical image segmentation network based on U-shaped structure, namely RAAU-Net model; as shown in the second aspect of the present application, Figure 1 As shown in the first aspect of the present application, the present application discloses a two-dimensional medical image segmentation network based on U-shaped structure, namely RAAU-Net model; as shown in the second aspect of the present application,

[0052] As shown in the first aspect of the present application, the present application discloses a two-dimensional medical image segmentation network based on U-shaped structure, namely RAAU-Net model; as shown in the second aspect of the present application, Figure 1 As shown in the first aspect of the present application, the present application discloses a two-dimensional medical image segmentation network based on U-shaped structure, namely RAAU-Net model; as shown in the second aspect of the present application, Figure 2 As shown in the first aspect of the present application, the present application discloses a two-dimensional medical image segmentation network based on U-shaped structure, namely RAAU-Net model; as shown in the second aspect of the present application,

[0053]

[0054] Wherein, Residual Basicblock, n represents the number of repetitions, i represents the number, MaxPool(·) represents the maximum pooling, Conv(7x7) represents the convolution operation with the convolution kernel size of 7x7.

[0055] The multi-receptive field extraction module (b), namely the IAR module, is used to extract semantic information of different receptive fields and generate more high-level feature maps, and the encoder-decoder structure is integrated. The output feature map of the encoding part can capture the context information and strengthen the use of global information.

[0056] Specifically, as shown in the first aspect of the present application, Figure 3As shown, the multi-receptive field extraction module, namely the IAR module, has a total of 4 cascaded branches, a single 3x3 convolution and an Avgpooling branch, and their receptive fields are all 3, which are mainly used to extract different feature information under the same receptive field. The other two branches are used to extract features of different scales. One branch uses two 3x3 convolutions to extract features. Although the receptive fields of the two are the same, the former uses fewer parameters. The other branch uses a dilated convolution with a dilation rate of 1, 2, and 3, which can increase the receptive field of the convolution without introducing additional parameters, and each pixel point in the receptive field is applied. The feature maps obtained by the 4 cascaded branches are added to the input feature map, which is similar to the shortcut mechanism in ResNet. The multi-receptive field extraction module can be represented by formula (2).

[0057]

[0058] output (·) The output feature map is represented by x, the input feature map is represented by F i (·) represents the output of the i-th branch.

[0059] The feature decoder module (c) decodes the obtained high-level feature map to restore the size of the initial image. The Squeeze-and-Excitation (SE) module used therein assigns corresponding weights to different channels for different contributions to the segmentation task, so that the network focuses on specific input parts.

[0060] Specifically, the feature decoder module has 5 layers. In the first 4 layers, each layer includes 1 TransposedConvolution, 1 Squeeze-and-Excitation (SE) module, and 2 3x3 convolutions. The 5th layer includes 1 TransposedConvolution, 1 3x3 convolution, and 1 1x1 convolution.

[0061] Attention mechanism is a resource allocation scheme that can handle more important information with limited computing resources and focus on unique parts in the case of limited computer capacity and a large amount of information. Figure 4 As shown, the Squeeze-and-Excitation (SE) module used in the application belongs to a channel attention module, which assigns corresponding weights to different channels for different contributions to the segmentation task, so that the network focuses on specific input parts. The feature map is input into the SE module, and a channel weight is obtained after global average pooling and a fully connected layer. The channel weight is multiplied by the feature map to obtain a new feature map.

[0062] The two-dimensional medical image segmentation method based on the U-shaped structure comprises the following steps:

[0063] Step S1: input the medical image to the feature encoder module, obtain rich and complex features of the input image by using the pre-trained ResNet-18, and accelerate the convergence speed of the network;

[0064] Step S2: input the feature map extracted by the feature encoder module to the multi-receptive field extraction module to extract semantic information of different receptive fields, and generate high-level feature map by using context information;

[0065] Step S3: the high-level feature map enters the feature decoder module after Squeeze-and-Excitation (SE); wherein, Squeeze-and-Excitation (SE) is used to assign different weights to different channels according to their different contributions to the segmentation task, so that the network focuses on specific parts of the input;

[0066] Step S4: the feature decoder outputs the final prediction segmentation map after decoding.

[0067] Further, step S1 includes the following steps:

[0068] Step S1-1: the input medical image is convolved by a convolution kernel with a size of 7x7 and a stride of 2, the image size is halved, and the channel number is 64;

[0069] Step S1-2: the output feature map of step S1-1 is subjected to maximum pooling and 2 layers of Residual BasicBlock, the image size is halved, and the channel number is 64;

[0070] Step S1-3: the output feature map of step S1-2 is subjected to 2 layers of Residual BasicBlock, the image size is halved, and the channel number is 128;

[0071] Step S1-4: the output feature map of step S1-3 is subjected to 2 layers of Residual BasicBlock, the image size is halved, and the channel number is 256;

[0072] Step S1-5: the output feature map of step S1-4 is subjected to 2 layers of Residual BasicBlock to obtain a feature map, and the channel number is 512.

[0073] Further, step S2 includes the following steps:

[0074] Step S2-1: the feature map extracted by the feature encoder module is subjected to convolution with a convolution kernel size of 3x3, BN layer and ReLU activation function to obtain the corresponding output;

[0075] Step S2-2: The extracted feature map of the feature encoder module is subjected to a convolution with a kernel size of 1x1 and two convolutions with a kernel size of 3x3, to obtain the corresponding output;

[0076] Step S2-3: The extracted feature map of the feature encoder module is subjected to three dilated convolutions with a dilation rate of 1, 2 and 3 respectively and a kernel size of 3x3, to obtain the corresponding output;

[0077] Step S2-4: The extracted feature map of the feature encoder module is subjected to 3x3 average pooling (Avgpooling) and a convolution with a kernel size of 1x1, to obtain the corresponding output;

[0078] Step S2-5: The extracted feature map of the feature encoder module is added element by element to the outputs of steps S2-1, S2-2, S2-3 and S2-4, to obtain a high-level feature map.

[0079] Further, step S3 includes the following steps:

[0080] Step S3-1: The deconvolution learns an adaptive mapping to restore features with more detailed information, and the feature map is upsampled by a deconvolution operation, with the width and height each being doubled;

[0081] Step S3-2: The feature map obtained by the deconvolution up-sampling is connected to the output of the corresponding convolution layer of the encoder through a jump connection;

[0082] Step S3-3: The connected feature map is input into a Squeeze-and-Excitation (SE) module, and a channel weight is obtained after global average pooling and a fully connected layer, and the channel weight is multiplied by the connected feature map to obtain a new feature map.

[0083] Step S3-4: The new feature map is then subjected to two convolutions;

[0084] Step S3-5: The steps S3-1, S3-2, S3-3 and S3-4 are repeated sequentially 4 times.

[0085] Effect description: The RAAU-Net proposed by the present application is applied to different two-dimensional medical image segmentation tasks. Taking the retinal nerve segmentation task as an example. In order to evaluate the model and expansion, the present application uses five commonly used segmentation task indicators. These indicators are pixel accuracy (PA), class pixel accuracy (CPA), IoU, recall (Recall), and Dice Score. As shown in Table 1, the present application compares the segmentation task indicator values with other segmentation methods to prove its effectiveness and advancement.

[0086] Table 1 Comparison of retinal nerve segmentation performance

[0087] Model PA (%) CPA (%) IoU (%) Recall (%) Dice (%) U-Net 95.19 82.34 68.20 79.88 81.09 SegNet 93.98 80.66 60.39 70.62 75.31 Deeplabv3 89.40 57.37 45.80 69.44 62.83 Mobilenetv3 91.21 68.87 46.18 58.36 63.18 MultiResUnet 95.04 83.15 66.97 77.49 80.22 R2UNet 94.72 80.47 65.72 78.19 79.31 Attention U-net 95.18 82.67 68.19 79.56 81.09 RAAU-Net 95.29 82.76 69.02 80.61 81.67 Best value 95.29 83.15 69.02 80.61 81.67

[0088] As Figure 5 shown, U-Net, MultiResUnet, RAAU-Net (the application) and the like are compared in qualitative results of retinal nerve segmentation using the DRIVE dataset. From left to right: Input Image, U-Net, MultiResUnet, RAAU-Net and Ground Truth.

[0089] The application adopts the above technical solutions, and the RAAU-Net model is composed of a feature encoder module, a multi-receptive field extraction module and a feature decoder module, wherein the multi-receptive field extraction module is proposed by the application and is composed of an Inception-ASPP-Residual (IAR) module. The module extracts semantic information of different receptive fields from the feature map output by the encoding part and generates more high-level feature maps. The application uses a pre-trained ResNet-18 to extract more rich and complex features, the multi-receptive field extraction module is used for integration of the encoder-decoder structure, the feature decoder module uses a Squeeze-and-Excitation (SE) attention mechanism to improve the decoding ability, and the whole model presents a U-shaped structure.

[0090] The application alleviates the loss of spatial information caused by the pooling operation by using the residual network module, and further improves the feature extraction capability of the network. A multi-receptive field extraction module is used to extract semantic information of different receptive fields, so as to capture context information, strengthen the use of global information, and strengthen the feature expression capability of the network. The Squeeze-and-Excitation (SE) attention mechanism is used to improve the decoding capability of the model, so that the model pays attention to the region of interest during the training process, and improves the segmentation effect of the model.

[0091] Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

Claims

1. A two-dimensional medical image segmentation network based on a U-shaped structure, characterized in that: It includes the following modules; Feature encoder module: The pre-trained ResNet-18 is used as the encoder to increase the depth of the network. At the same time, the branch adopts the shortcut mechanism to avoid gradient vanishing and accelerate network convergence. Downsampling uses max pooling and strided convolution to reduce the image size. The multi-receptive-field extraction module is used to extract semantic information from different receptive fields and generate more high-level feature maps. It integrates an encoder-decoder structure and can capture contextual information in the output feature map of the encoding part, enhancing the utilization of global information. The module includes four cascaded branches: a single 3x3 convolutional branch and an average pooling branch are used to extract different feature information within the same receptive field, each with a receptive field of 3. The other two branches are used to extract features at different scales. One branch uses two 3x3 convolutions to extract features; the other branch uses dilated convolutions with dilation rates of 1, 2, and 3 to increase the receptive field without introducing additional parameters and applies the dilation to every pixel within the receptive field. The feature maps obtained from the four cascaded branches are added to the input feature map. The formula for expressing the multi-receptor field extraction module is as follows: output (·) Let F represent the output feature map, x represent the input feature map, and F represent the output feature map. i (·) represents the output of the i-th branch; Feature decoder module: decodes the obtained high-level feature map and restores it to the size of the initial image; The SE module assigns weights to different channels based on their contribution to the segmentation task, allowing the network to focus on specific input parts. The feature decoder module has 5 layers. The first 4 layers each consist of 1 deconvolution, 1 SE module, and 2 3x3 convolutions. The 5th layer consists of 1 deconvolution, 1 3x3 convolution, and 1 1x1 convolution. The feature map is input into the SE module, and after global average pooling and a fully connected layer, a channel weight is obtained. The channel weight is multiplied by the feature map to obtain a new feature map.

2. The two-dimensional medical image segmentation network based on a U-shaped structure according to claim 1, characterized in that: The feature encoder module initially uses strided convolutions with a stride of 2 to reduce the image size, then uses max pooling to further reduce the image size, followed by another strided convolution with a stride of 2 to reduce the image size. After two convolutions, the input feature map becomes the output feature map. The adjusted input feature map and the output feature map are then added element-wise to increase the feature information in each dimension. Finally, the ReLU activation function is used to obtain the final output. The formula for the feature encoder module is as follows: in, This represents a Residual Basicblock, where n represents the number of repetitions, i represents the number of elements, and MaxPool(·) represents max pooling. This indicates a convolution operation with a kernel size of 7x7.

3. A two-dimensional medical image segmentation method based on a U-shaped structure, employing the two-dimensional medical image segmentation network based on a U-shaped structure as described in claim 1, characterized in that: The method includes the following steps: Step S1: Input the medical image into the feature encoder module and use the pre-trained ResNet-18 to obtain rich and complex features of the input image, thereby accelerating the convergence speed of the network. Step S2: Input the feature map extracted by the feature encoder module into the multi-receptive field extraction module to extract semantic information of different receptive fields, and generate a high-level feature map using contextual information; Step S3: The high-level feature map enters the feature decoder module after passing through the SE module; the SE module assigns different weights to different channels based on their different contributions to the segmentation task, so that the network focuses on specific parts of the input. Step S4: The feature decoder outputs the final predicted segmentation map after decoding.

4. The two-dimensional medical image segmentation method based on a U-shaped structure according to claim 3, characterized in that: Step S1 includes the following steps: Step S1-1: The input medical image is subjected to strided convolution with a kernel size of 7x7 and a stride of 2, resulting in a half-size image and 64 channels. Step S1-2: The output feature map from step S1-1 is subjected to max pooling and two layers of Residual Basic Block to halve the image size, resulting in 64 channels; Step S1-3: The output feature map from step S1-2 is halved in size after passing through two layers of Residual BasicBlock, resulting in 128 channels; Step S1-4: The output feature map from step S1-3 is halved in size after passing through two layers of Residual BasicBlock, resulting in 256 channels; Step S1-5: The output feature map from step S1-4 is processed by two layers of Residual Basic Blocks, reducing the image size by half to obtain a feature map with 512 channels.

5. The two-dimensional medical image segmentation method based on a U-shaped structure according to claim 3, characterized in that: Step S2 includes the following steps: Step S2-1: The feature map extracted by the feature encoder module is processed by a 3x3 convolution kernel, a BN layer, and a ReLU activation function to obtain the corresponding output; Step S2-2: The feature map extracted by the feature encoder module is processed by a convolution with a kernel size of 1x1 and two convolutions with a kernel size of 3x3 to obtain the corresponding output; Step S2-3: The feature map extracted by the feature encoder module is subjected to three dilated convolutions with dilation rates of 1, 2 and 3 and a kernel size of 3x3 to obtain the corresponding output; Step S2-4: The feature map extracted by the feature encoder module is subjected to 3x3 average pooling (Avgpooling) and convolution with a kernel size of 1x1 to obtain the corresponding output; Step S2-5: The feature map extracted by the feature encoder module is added element by element to the outputs of steps S2-1, S2-2, S2-3, and S2-4 to obtain the advanced feature map.

6. The two-dimensional medical image segmentation method based on a U-shaped structure according to claim 3, characterized in that: Step S3 includes the following steps: Step S3-1: Deconvolution learns adaptive mapping to recover features with more detailed information. The feature map is upsampled by deconvolution, doubling its width and height. Step S3-2: The feature map obtained by deconvolution upsampling is connected to the output of the corresponding convolutional layer of the encoder through skip connections; Step S3-3: The connected feature map is input into the SE module, and after global average pooling and fully connected layers, a channel weight is obtained. The channel weight is multiplied by the connected feature map to obtain a new feature map; Step S3-4: The new feature map is then convolved twice more; Step S3-5: Repeat steps S3-1, S3-2, S3-3, and S3-4 in sequence 4 times.

Images